ChatGPT's New Memory Actually Remembers You Now. That's the Feature and the Problem.

Dreaming V3 attacks exactly that. The system now prioritizes recent context over old, and updates memories as time passes on its own. OpenAI's example is the system revising "You are going to Singapore in July" into "You went to Singapore in July 2026" after the trip ends. There's a memory summary you can open, review, and edit anytime, and you can tell it when to apply which preferences.

The benchmark deltas are unusually large for a feature update: preference adherence up from 55.3 to 71.3 percent, and accuracy over time up from 52.2 to 75.1 percent. In normal use that second number is the one you feel. The system's mental model of you stops decaying.

In my own testing the difference showed up quickly. It correctly stopped suggesting a framework I'd told it months ago I had migrated away from, something the old memory got wrong constantly.

The part I keep turning over

Memories now update automatically, with ChatGPT deciding which details are important enough to keep building on. And as TechTimes noted, the rewritten personalization engine limits the audit trail compared to the old discrete-memory list.

The old system was dumb but legible: a flat list of remembered facts you could read line by line and delete one at a time. Dreaming V3 gives you a summary, a written digest of what the model believes about you, rather than the underlying ledger of what was stored, when, and from which conversation. You can edit the digest. You can't fully inspect the bookkeeping behind it.

For a feature whose entire job is building a persistent model of your life, the editability is genuinely good and the reduced inspectability is genuinely not. A summary you can edit is not the same thing as a record you can audit. Anyone who has ever filed an expense report knows the difference.

Should you turn it on?

If you already use ChatGPT as a daily tool, yes, probably. The quality jump is real, the Singapore-style staleness problem was the feature's biggest flaw, and the controls that do exist, review, edit, and per-preference instructions, are better than what they replaced.

If your conversations touch things you'd rather not have synthesized into a profile, health, finances, work matters under NDA, this is a good week to open the memory summary, read what's actually in there, and prune. The system deciding "what's most important" about you is exactly the moment to check its judgment.

Go read your summary

Dreaming V3 is what memory should have been from the start, and the 20-point accuracy gains will make assistants from every vendor chase it within a quarter. Just go read your memory summary this week. Not because anything sinister is in it, but because from now on, what that page says is what the machine thinks you are, and the receipts behind it are getting harder to see.